You are on page 1of 15

216

International Journal of Management

Vol. 24 No. 2

June 2007

A Comparative Test of Two Employee Turnover Prediction Models
Wei-Chiang Hong Oriental Institute of Technology, Taiwan Ruey-Ming Chao Oriental Institute of Technology, Taiwan Accurate employee turnover prediction models are critical in the early detection of unanticipated turnover, giving managers sufficient time to deal with turnover related management issues. The logit and probit models have been successfully applied to solve nonlinear classifying and regression problems. However, the feasibility of applying them to voluntary turnover prediction still has not been sufficiently explored. Consequently, a numerical example involving voluntary turnover data fmm a motor marketing enterprise in central Taiwan is employed, and the empirical results reveal that the proposed models have high prediction capabilities. Accordingly, these two models also provide a promising alternative for predicting employee turnover in human resource management.

1 Introduction
1.1 Employee turnover An increasing number of models that accurately predict employee turnover provide early detection of unanticipated turnover, particularly from high performance employees. Meanwhile, accurate prediction models also leave managers with adequate time to deal with turnover related management issues (Morrow. McElroy. & Laczniak, 1999). Generally, two forms of employee turnover exist, involuntary turnover and voluntary turnover. Involuntary turnover is frequently delined as movements across organizational boundaries, over which the employee is only slightly affected (Price, 1977), the representation forms of withdrawal from an organization often categorized with strategy (e.g. downsizing), dismissal (e.g. firing), or policy (e.g. compulsory retirement) (Campion, 1991). In contrast, voluntary turnover is defined as movements across organizational boundaries, over which the employee is heavily affected (Price, 1977), the representation forms of withdrawal from an organization often categorized with absenteeism and lateness (Campion, 1991). In accordance with the above definitions, involuntary turnover is most often initiated by organizational changes and institutional constraints, and it is easy to understand the scope and the affections for the organization. However, employers find it especially difficult to precisely predict the withdrawal forms, the affection and the timing t)f voluntary turnover in an organization, such as the number of individuals leaving, which individuals are leaving, and how individuals are leaving. In other words, the employers have to understand the damages resulting from high performance employees leaving, and the benefits resulting from poor performance employees leaving (Dalton & Todor, 1982; Simons & Hinkin, 2001) as well as the methods of withdrawal used by

International Joumal of Management

Vol. 24 No. 2

June 2007

217

departing employees (Morrow et a!., 1999). Tberefore, the voluntary turnover modeling methodology and novel prediction approach have attracted considerable attention during recent decades. Mobley (1977)firstproposed structural models for representing employee turnover: job dissatisfaction — thinking of quitting — intention to search (for a new > > job)—> intention to quit-> (voluntary) turnover. Based on this brief mcxlel of voluntary tumover, job satisfaction is an indirect link in initialing actual voluntary turnover (Muller, Boyer, Price, & Iverson, 1994; van Breukelen, van der Vlist, & Steensma, 2004). Researchers also added organizational commitment as an intervening variable to explain the outcomes, stay intentions or tumover (Inverson, 1992; Muller, Wallace, & Price, 1992; Price, 1986). In addition, numerous empirical research studies have suggested that greater organizational commitments imply greater intent to stay, i.e., lower voluntary turnover (Lincoln & Kalleberg, 1996; Muller et a!., 1992; Price, 1986; Somers, 1995). Job performance bas become an increasingly popular influence on decisions related to employee turnover, due to its functional indication of withdrawal process for understanding the management status of an organization. For example, losing high performance employees implies high human resource costs and implicates organizational operations. The relationships between performance and turnover can be briefly divided into four categories, a positive relationsbip, a negative relationship, no relationship and a non-linear (curvilinear) relationship. (Jackofsky, 1986). Birnbaum and Somers (1993) suggested that no relationship exists between employee performance and turnover. Furthermore, Lazarsfeld and Theilens (1958), and Schwab (1991) presented evidence of a positive relationship between performance and turnover. They indicated higher turnover among high performance employees. However, Steers and Mowday (198!) indicated that high job performance heightened both expected and actual organizational rewards, while low job performance caused low attitudes regarding the intrinsic worth of a job. Similarly, Jackofsky (1984) suggested that iow job performance is asscKiated with high voluntary turnover, due to employees realizing tbat tbeir employment is in peril. Therefore, tumover displays a curvilinear relationship with low turnover being associated with medium job performance, while higb turnover is associated witb extremely good or bad job performance. Lance (1988) and Trevor, Gerbart, and Boundreau (1997) applied the curvilinear relationship to conduct a moderator of employee turnover via salary growtb and promotions. Tbe negative relationship between performance and turnover appears to be the major conclusive finding, indicating that high performance employees would be less likely to leave than lower performance ones. McEvoy and Cacsio (1987), Vecchio and Norris (1996), and Williams and Livingstone (1994) found the correlations between the two (high performance employees and lower performance ones) are -0.19, -0.28 and -0.16, respectively. Moreover, Morrow et al. (1999) also demonstrated that a statistically significant negative relationship exists between the two ones. The profound recommendations of negative relationship from Williams and Livingstone (1994) are Ihal while those organizational reward systems were contingent on or tied to performance, the employees with poor performance would tend to leave, particularly in organization marketing departments.

218

International Joumal of Management

Vol. 24 No. 2

June 2007

1.2 Binary choice approaches to voluntary tumover prediction As mentioned previously, employee turnover should be viewed as the initiated discrete choice of each employee. Therefore, the voluntary turnover prediction model could be suitable for discrete choice models. Generally, discrete choice models include the initiated choice (turnover choice) model and the passive choice (under objective pressure) model. Discrete choice models are used to predict choices of alternatives among numerous discrete altematives. The chosen alternatives are not only based on tbe cbaracteristies of cboosers (individual performance) but also the attributes of each alternative. In this study, multinomial choice only implies tending to leave the organization or remain there, i.e., the choice is binomial. Discrete choice models are now widely applied in social science, particularly with the increasing apphcation of McFadden's (1973) multinomial (or binomial) logistic regression model (namely logit model). Kim and Arbel (1998) applied the logit model to predict merger target selection fora hospitality company, and claimed tbat tbe logit model can be used as a supplementary decision-supporting tool. Nassimberni (2001) proposed using the logit model to predict the altitude towards exporting of small manufacturing firms. The studies indicated that tbe logit model provides a robust approach which showed that exports are strictly linked to firms' ability to innovate the products and to valid inter-organizational relations rather than technological profiles. Similarly, Kim and Kwon (2003) applied the logit model to detect Korean mobile telephone market segmentation, and indicated that the logit modei successfully demonstrated that the consumers preferred carriers dealing with other things even although they have a large number of subscribers. Meanwhile, Kim and Kwon also reported an approach for increasing consumer sources, such as intra-network call discounts and quality signaling effects. Additionally, Tseng and Yu (2005) proposed a MNL model combined with fuzzy integral technology to partition tbe Taiwanese Internet telephone market into four segments (market share), IP devices (7%), computers (5%), IP cards (22%), and traditional phones (66%). Recently, logit models have also been applied to forecasting choice between two discrete alternatives (fail or non-fail), such as forecasting growth of market share (Agrawal & Schoriing, 1996), bankruptcy forecasting (Tseng & Lin, 2005) and business cycle (Layton & Katsuura, 2001). However, logit models have the property of the relative probabilities of eacb alternative being independent in terms of tbe presence or characteristics over all other alternatives. This properly, known as the independence of irrelevant alternatives (IIA), indicates that introducing or improving any alternative will equally and proportionally impact the probability of each otber alternative (Stopber, Meyburg, & Brg, 1981). Tbis representation of choice bebavior is impractical and results in biased estimates and inconect predictions: for example, any improvement of tumover altematives (leave or stay) never has equal impact probability with each otber. To relax the IIA property restriction of the logit model, the multinomial (binomial) probability regression model (namely probit model) allows a completely free correlation structure among the discrete choice alternatives. Similarly, probit models have been employed in binary choice

International Journal of Management

Vol. 24 No. 2

June 2007

219

prediction: Dow and Endersby (2004) compared the mtiltinomial logit model (MNL) and the multinomial probit model (MNP) for analyzing voting choice in the United States and France, and found that the simpler logit model i.s frequently preferable to tbe more complex pro/j//regarding multi-party elections. Therefore, the/o^'//and/?ra/?f7 models are highly suitable for application to employee voluntary turnover qualitative choices. This investigation employs the negative relationship between performance and voluntary turnover to construct the employee leaving prediction model. Two conventionally adopted statistical forecasting approaches, tbe basic (binomial) logit and basic (binomial) probit models, are employed to compare their prediction performance. The rest of this paper is organized as follows. Section 2 introduces the proposed logit and probit models. Section 3 then presents a numerical example involving a motor marketing enterprise in Taiwan to demonstrate the predictive capacity of the proposed model. Conclusions are drawn in section 4. .

2 Model specifications 2.1 Logit model
This study assumes that the state S., for each employee observation, appears absolute certainty in the turnover prediction models. Thus, let S^ = 1, when an employee has left the organization, i.e., voluntary turnover; in contrast, 5.^0, when an employee remains in the organization, i.e., non-turnover. The turnover variable S is modeled directly as the dependent variable in a model using the logistic function, represented as Eqs. (1) and (2), respectively, exp(/Jc) 1 P(S.= ,turnover) = ^ ^' = -— (1) P(S- = 0, Qon-tumover) = -; -^— (2) ', I I + exp(/Jx-.) where Jtj denotes the explanatory variable for the turnover variable., in this investigation X- represents the job-performance of eacb employee /; ^are the variable coefficients of the logit model and would be estimated. Equations (I) and (2) are then estimated by maximum likelihood estimation (MLE), with the turnover/non-turnover information used as the training data set. The values of variable coefficients, /3., are calculated for each technique as Ihe log odds ratio using the estimated probability of the teehnique. After determining the values of /?., estimation is easily performed using any software with a loglt regression routine. Finally, the prediction values of turnover variable S obtained from this model forecast 5. being either voluntary turnover (5^. = 1) or non-turnover (5, = 0). 2.2 Probit model As in the logit model, assume that the state S-, for each etnployee observation, appears with absolute certainty in the turnover prediction models. However, the logit and probit models differ in that, rather than the logit function, a cumulative standard

220

International Journal of Management

Vol. 24 No. 2

June 2007

normal distribution functional form (i.e., with unit variance and zero mean value) is employed in this section. Therefore, when an employee voluntarily decides to leave, let S- = 1; otherwise, in contrast, 5. = 0, when an employee decides to remain, let S- = 0. The turnover variable S is modeled directly as the dependent variable in a model of the form as Eq.(3), (3) e'''"dx (4)

where O() denotes the cumulative standard normal distribution function, and is defined as Eq.(4); jf. is also represented as the job-performance of each employee / for the turnover variable; ^are the estimated variable coefficients o(\heprobit model. Equation (3) is once again estimated by MLE via the training data set of lurnover/nonturnover. The values of variables coefficients. /3 . are also calculated for eacb technique as the log odds ratio using the estimated probability of that teehnique. The estimation is easily performed using any software with a probit regression routine. Finally, the prediction values of turnover variable S^ can also be calculated from Ihis probit model, being eitber voluntary turnover (S = 1) or non-turnover ( 5 . - 0). Both the logit and probit models guarantee that the estimated probabilities are between 0 and 1. However, no compelling theoretical reasons exist for either model being superior, or more suitable in specific situations. Further details can be found in Greene (2003). 2.3 Measure indice.s of prediction performance Although one model can rarely be the optimum choice for a given set of employee turnover data, the accuracy of the prediction model needs to be assessed so that the model that generally works best and produces the smallest error can be identified. There are a number of methods of calculating, measuring and interpreting errors, for example Cox and Snell R^ (Cox & Snell, 1989). Nagelkerke R^^ (Nagelkerke, 1991), McFadden R~ (McFadden, 1973), and the classification table and model Chi-square test (Greene, 2003). However, because this study focuses on predicting whether an individual will remain or leave an organization based on their job-performance, the success rate of correct prediction (based on a classification table) should be the primary consideration, rather than evaluating the explanatory capability of the proposed model. Diebold and Rudebusch (1989) proposed an effective method of quantifying the above. This method is widely known as the Quadratic Probability Score (QPS), defined in Eq. (5), QPS = x^(P -D.)^ (5)

where T denotes Ihe number of prediction periods; D takes the value 1 in Ihe event of voluntary turnover and 0 otherwise; P is the model derived probability for employee /. If the value of QPS is near 0. it implies an accurate model.

International Journal of Management

Vol. 24 No. 2

June 2007

221

This study employs the classitication table and QPS to examine how the prediction model performs. Meanwhile, the proposed models are implemented by statistical software, SPSS for Windows 10.0 version. Some useful statistics, Cox and Snell R , Nagelkerke R~, MeFadden R and the Chi-square test, are also introduced in the numerical example for reference.

3 Numerical example
3.1 Data set and research structure The sample was comprised of 150 professional employees drawn from a motor marketing company located in the central Taiwan, which has operated since 1992, and has an annual business volume of 300 million NTD and over 2(K) marketing specialists. Due to missing data (item non-response), Ihe useabie sample size was 132. All participants had direct responsibility for motor marketing, motor maintenance, and held the minimum of an associate degree for sales. Attitudinal data were gathered from a questionnaire survey that was distributed and collected on Ihe intranet web-site. Respondents were given the opportunity to complete the questionnaire during their work time and were asked to provide their employee personnel numbers to allow access to their personnel files. Confidcniialiiy of all information collected was assured and maintained. The response rate to the survey was 88 percent. The sample was 83 percent male, 3 percent Aborigine and 15 per cent Hakka, and had a mean age of 31.7 years old. In this investigation, the data set is used to demonstrate the forecasting accuracy of the proposed models. The data contained a total of 132 marketing specialists and their voluntary turnover status for 2003. The variables included in this study are composed of two distinct subsets. The first subset, represents the antecedents of turnover from theory and researchfindings:job-performance (mainly assessed by motor sales volumes in 2003), organizational commitment and job withdrawal intentions (Meyer & Allen, 1990; Mobley, 1982; Mowday, Steers, & Porter, 1982). The second subset of variables is composed of non-work influences on turnover and includes: age, number of dependent children, and race/ethnicity (Aborigine, Hakka, and Han). Although turnover is usually influenced by work and non-work factors, turnover research has concentrated primarily on work-related variables (Mobley, 1982; Mowday etal., 1982). However, it is interesting to see that some non-work antecedents of turnover were consistent with a time-based variable which had some relationship to work and family. For example, age is a timebased variable that affects one's job mobility and one's stage in the personal and family life-cycles. Similarly, number of dependent children is a non-work influence of turnover that is linked to life-stage and job mobility. Finally, race in Taiwan has been shown to infiuence expectations about job-performance. Due to the stigma of being a minority. Aborigine (2 per cent of total population, see Taiwan Government Information Office, 2005) and Hakka (28 per cent, see Taiwan Government Infonnation Office, 2005) in Taiwan often have diminished expectations about their job-performance. During several thousand years, these Aborigines and Hakka lived in isolation in the plains, coastal, and mountain areas of Taiwan. Because of subsequent migration, ihe Aborigines and

222

International Joumal of Management

Vol. 24 No. 2

June 2007

Hakka in the coast and plains were progressively assimilated by immigrant groups, and now most identifiable aboriginal tribes are limited to less accessible regions in the mountains. Aborigines and Hakka sometimes viewed by tbe Taiwanese mainstream culture as lacking in loyalty, passion, and responsible. Therefore, race can be viewed as a non-work influence of turnover tied to relationsbips between race and turnover. Tbe research structure of this investigation is shown in Fig. I. 3.2 Prediction procedures The data set is divided into two parts, the modeling data set (containing employees one to 100) and the testing data set (containing employees 101 to 132). The modeling data set is used to test the logit and probit models. The testing data set is not used for either model building or selection, and is used for estimating model performance when applied to future data. Table I lists the different data sets for this example. Since the logit and probit prediction models include the explanatory coefficient variables (constant item and job-performance), /3 , requiring estimation via the MLE procedure, the Maximum likelihood estimation procedure can be employed to determine the free parameters for the logit and probit models. Table 2 lists tbe results of these coefficients for each proposed model. For the logit model, the explanatory capability exceeds 50% (the Nagelkerke R^ is 0.529); the two estimated coefficients 0Q and /?, are also statistically significant, and have values of -0.206 and 2.343 respectively. For

Figure 1 Research structure of voluntary turnover
Age Annua motor j sales volumes 1
< 1

Gender Job-performance Race

Organization committee

International Joumal of Management

Vol. 24 No. 2

June 2(M)7

223

tbe probit model, tbe explanatory capability also exceeds 50% (the Nagelkerke's R' is 0.521); the two estimated coefficients ^^ and ^j are also statistically significant, and bave values -0.112 and 1.361 respectively. Table 3 summarizes the prediction accuracy during tbe training stage for each proposed model. For the logit model, the success rate for predicting individual employee turnover is 87.3%. In contrast, the success rate for predicting non-turnover is only 57.8%, and the total rate of correct prediction 74%. The QPS is 0.1829. For the prohit model, the rate of correctly predicting individual employee tumover is 89.1%, however, the rate of correctly predicting non-turnover is only 51.1%. Finally, the QPS is 0.2976.
I

3.3 Prediction results Table 4 compares the prediction results obtained using the proposed models in the testing stage. For the logit and probit models, tbe success rates for prediction of employee tumover hotb reacbed 95.2%, wbicb tbe success rates for predicting nonturnover reached only 27.3%; thus tbe total success rate for predicting was 71.9%. Tbe QPS of applying the logit and probit models was 0.1240 and 0.1689 respectively.

Table 1 Various data sets for the logit model and probit models
Logit / Probit models Data periods Training data periods Testing data periods 1st-132nd Ist-lOOth iOlst-132nd

Table 2 Coefficients of explanatory variables for logit and probit models
Logii model Explanatory Estimated variables coefficients Stand error 0.041 0.437 0.000** 0.387 Probil model p-value Explanatory Estimated variables coefficients Stand error 0.020 0.228 p-value

(A,and^,)
Constant Jobperformance -0.206 2.343 0.000** 0,000** Constant Jobperformance

(^oand/?,)
-0.112 1.361 0.000** 0.000**

Model Chi-square p-value (d.f.=l) • Cox and Snell/;-

Model Chi-square p- 0.000** value (d-f.=l) Cox and Snell/?^ Nagelkerke/;^ McFadden;;0.382 0.521 0.365

Nagelkerke R^^ 0.529 McFadden R~ 0.372

224

International Journal of Management

Vol. 24 No. 2

June 2(M)7

Table 3 The prediction accuracy in the training stage for logit and probit models
Logit model Actual turnover (5,=l) Predict as "turnover" Predict as"nontu mover" Total Success rate for predi cling Actual nontumover iS,=O) 19 Total Probit model
AclULlI

turnover

(S=\)
67 Predict as "turnover" Predict as"nonturnover" Total Success rate for predicting

Actual nonturnover (S^=0) 22

Total

48

49

71

7 ?5 87.3% 0.1829

26 45 57.8%

33 100 74.0%

6 55
89.1% 0.2976

23

29 100
72.0%

45
51.1%

OPS

OPS

Table 4 The prediction accuracy in the testing stage for logit and probit models
Logit model Actual turnover (5,=1) Predict as "turnover" Predict as "nontu mover" Total Success rate for predicting Actual nontumover (S,-=0) Total Probil model Actual tumover (5,=!) Predict as "tumover" Predict as "nontu mover" Total Success rate for predicting OPS Actual nonturnover (S,=0) 8 3 11 27.3% Total

19
2 21 90.5% 0.1240

8 3
11 27.3%

27 5 32 71.9%

19 2 21 90.5% 0.1689

27 5 32 71.9%

OPS

International Journal of Managemenl

Vol. 24 No. 2

June 2(M)7

225

From a statistical perspective, the model Chi-squarc te.sts are signilicant for both these two models, but the logit model is superior to the probit model in terms ot explanatory capability (the value.s of Cox and Snell /?', Nagelkerke v and McFadden R' are slightly better than {)^'S-probit model). However, during the testing stage., the classilicalion test (Table 4) indicates that the prediction performance of the logit model equals that of the probit model. Therefore, the Quadratic Probability Score (QPS), as given by Eq.(4). is employed to demonstrate the prediction accuracy of these two models. As mentioned above, the minimum values of QPS indicate that the deviations between aetual and forecast values are very small. The comparison results demonstrate that the prediction accuracy ofthe logit model is superior to that o{ ihc probit model (in bolh Ihe training and testing stages).

I

4 Discussions and Conclusions

Sinee aceurately predicting employee turnover is essential in the eariy detection of unanticipated turnover of an organization, particularly for high performance employees, on appropriate and efficient prediction model must be adapted to conduel turnover forecasting in advance. Turnover research should move in new directions based on new assumptions and methodologies. This would raise new issues and problems (such as involves neural networks, support vector machines to conduct classification problem, i.e., stayer or leaver). This investigation applies logit and probit models to predict employee voluntary turnover. Experimental results indicate that the proposed technique provides an effective alternative for predicting employee voluntary turnover for a non-linearly combined variable, namely job-performance. At present too many studies of employee turnover use simple job-performance variables (age, race, gender, organization committee, and annual motor sales volumes) for us loreaeh any definitive conclusions about the utility of job-performance based prediction models of studying employee turnover. Our findings provide a convenient research context to aecurately predict the willing-decision of high perfomiance employees in the future. Interestingly, our model indicate.s that raee/ethnieity is an important factor in job-performance which affects withdrawal behavior. For example, in the testing stage, it was found that, in Table 4, among the 32 employees there wasonly one Aborigine (located in the grid with "predict as non-turnover" and "aetual turnover"), and 10 Hakka (kx;ated in the grid with "predict as tumover" and "actual non-tumover"). This implies that Aborigine and Hakka do not behave in such a way that the rules from mtxieling (logit or probit models) could be applied to them. The present study offers a different perspective, i.e., race as a factor affecting the perceived value of one's investments in an organization, which is consistent with eommitment theory (Meyer & Allen, 1984) and theory and research on the topic of demographic diversity in organizations (Cox. 1994). Similarly, our model also provides evidence that gender is a secondary factor in affecting withdrawal behavior. Females in Taiwan are traditionally expected to pay more attention to their family, even now that Taiwan is a developed country. In table 4, it was found that there ten females (located in the grid with "predict as turnover" and "aetual turnover") determined the turnover results. Females in Taiwan appear to withdraw depending on the number of children and

226

International Journal of Management

Vol. 24 No. 2

June 2001

the lime required to take care of their children, such that modeling could easily explain their decision making. It suggests that gender or the numher of children are factors affecting one's turnover decision in an organization. Finally, turnover might not always he the result of a rational, ordered sequence hased on careful consideration of allernalivc opportunities (Lee & Mitchell, 1994), but rather might he driven hy work attitudes such as organizational cnmmilment. More generally, if the organization is designed for high performance, then one would expect turnover hehavior to occur more rapidly and/or without full consideration of its immediate personal consequences. The perception of the organization committee and individual expectationa are the most important factors to understand in order to explain, and to predict the withdrawal hehaviors of high performance employees. In table 4, it was also found that twelve employees (located in the grid with "predict as turnover" and "actual turnover") determined turnover perhaps because this motor marketing company violated some aspects of the performance honus policy during the hard-market exploration-period in 2(X)3. Almost all of the Top 20 sales managers (in this case is twelve employees including three regional managers, four project managers, and five senior specialists) decided to leave the organization after Lunar New Year. This case reveals that an organization committee mainly affects high performance employees' investments in an organization, even if the committee was more than they expected. If this kind of committee fails, they will find it difficult to believe any important committee of this organization. Maybe job withdrawal is the best policy for them under this kind of situation. In this sense, commitment reflects prior behavior, while joh withdrawal intentions are future oriented. The previous discussion is based on rigid modelsmodeling processes. However, turnover behavior is not always as rational or as logical as results from turnover research would have us believe; turnover might not always he the result of a rational decision-making process characteristic of economic models of human behavior. This raises the issue of whether the most widely accepted methods for studying turnover are sensitive enough to detect deviations from the conceptual model that they are intended to test. Future research could include other high independent variables, including job-satisfaction, the gap between organization commitment and individual expectation (Allen & Griffeth, 1999; Currivan, 1999; Sager, Griffeth. & Hom, 1998). or the abnormal absenteeism of some employees (Morrow et al., 1999), into the MNL and MNP models to enhance their predictive power.

References
Agrawal, D., & Schoriing, C. (1996) Market share forecasting: An empirical comparison of artificial neural networks and multinomial logit model. Journal of Retailing, 72: 383-407. Allen, D.G., & Griffeth, R.W. (1999) Job performance and turnover: A review ami integrative multi-route model. Human Resource Management Review, 9(4): 525-548. Birnbaum, D., & Somers, M.J. (1993) Fitting job performance into the turnover model: An examination of the form of the job performance^—Turnover relationship and a path model. Joumal of Management, 19: 1-11.

International Joumal of Management

Vol. 24 No. 2

June 2007

227

Campion, M.A. (1991) Meaning and measurement of turnover: Comparison of alternative measures and recommendations for research. Journal of Applied Psychological. 76(2): 199-212. Cox, D.R., & Snell, E.J. (1989). The analysis of binary data. London: Champmanand Hall. Cox, T. (1994). Cultural diversity in organizations: Theory, research and practice. San Francisco, CA: Berett-Koehler Puhlishers. : . Currivan, D.B. (1999) The causal order of joh satisfaction and organizational commitment in models of employee turnover. Human Resource Management Review, 9(4): 495-524. Dalton, D.R., & Todor, W.D. (1982) Turnover overstated: A Functional taxonomy. Academy of Management Review, 7: 117-123. Diebold, F.X., & Rudebusch, R. (1989) Scoring the leading indicators. Journal of Business, 62: 369-402. Dow, J.K., & Endershy, J.W. (2004) Multinomial prohit and multinomial logit: a comparison ofchoice models for voting research. Electoral Studies, 23: 107>122. Greene, W.H. (2003). Econometric analysis. New Jersey: Prentice Hall. Inverson, R.D. (1992). Employee intent to stay: An empirical test of a revision of the Price and Muller model. Iowa City: The University of Iowa (unpublished doctoral dissertation). Jackofsky, E. (1984) Turnover and job performance: An integrated process model. Academy of Management Review, 9: 74-83. Jackofsky, E.F., Ferris, K.R., & Breckenridge, B.G. (1986) Evidence for a curvilinear relationship between joh performance and turnover. Journal of Management, 12: 105III. Kim, H.S., & Kwon, N. (2003) The advantage of network size in acquiring new subscribers: a conditional logit analysis of the Korean mobile telephony market. Information Economics and Policy, 15: 17-33. Kim, W.G., & Arbel, A. (1998) Predicting merger targets of hospitality firms (a logit mode]). Hospitality Management, 17: 303-318. Lance, C. (1988) Job performance as a moderator of the satisfaction-turnover intention relation: An empirical contrast of two perspectives. Joumal of Organizational Behavior, 9:271-280. Layton, A.P., & Katsuura, M. (2001) Comparison of regime switching, probit and logit models in dating and forecasting US business cycles. International Journal of Forecasting, 17:403-417. Lazarsfeld, P., & Theilens. W. (1958). The academic mind: Social scientists in a time of crisis. Glenview, IL: The Free Press.

228

International Joumal of Management

VoL 24 No. 2

June 2007

Lee, T. & Mitchell, T. (1994) An alternative approach: the unfolding model of voluntary employee turnover. Academy of Management Review, 19: 51-89. Lincoln, J.R., & Kalleberg, A.L. (1996) Commitment, quits and work organization: A study of U.S. and Japanese plants. Industrial and Labor Relations Review, 50: 738-760. McEvoy, G.M., & Cacsio, W.F. (1987) Do good or performers leave? A meta-analysis of the relationship between performance and tumover. Academy of Management Joumal, 30: 744-762. McFadden, D. (1973). Conditional logit analysis of qualitative choice behavior. In P. Zorcmmhka (Eds.), Frontiers in econometrics. New York: Academic Press. Meyer, J., & Allen, N. (1984) Testing the side-bet theory of organizational commitment: reexamination of the continuance and affective states. Joumal of Applied Psychology, 69: 372-378. Mobley, W.H. (1977) Intermediate linkages in the relationship between job satisfaction and employee turnover. Joumal of AppUed Psychology, 63: 237-240. Mobley,W.H. (1982). Employee turnover: Causes, consequences and control. Reading, MA: Addison-Wesley. Morrow, P.C., McEIroy, J.C, & Laczniak, K.S. (1999) Using absenteeism and performance to predict employee turnover: early detection through company records. Joumal of Vocational Behavior, 55: 358-374. Mowday, R., Steers, R., & Porter, L. (1982). Employee-organization linkages. New York: Academic Press. Muller, C.W., Wallace, J.E., & Price, J.L. (1992) Employee commitment: Resolving some issues. Work and Occupations, 19: 211-236. Muller, C.W., Boyer, E.M., Price, J.L., & Iverson, R.D. (1994) Employee attachment and noncorcive conditions of work: The case of dental hygenists. Work and Occupations, 21:179-212. Nagelkerke, N.J.D. (1991) A note on a general definition of the coefficient of determination. Biometrica, 7: 691-692. Nassimbeni, G. (2001) Technology, innovation capacity, and the export attitude of small manufacturing firms; a logit/tobit model. Research Policy, 30: 245-262. Price, J.L. (1977). The study of tumover. Ames: Iowa State University Press. Price, J.L., & Muller, C.W. (1986). Handbook of organizational Cambridge, MA: Ballinger. measurement.

Sager, J.K., Griffeth, R.W., & Hom, RW. (1998) A comparison of structural models representing turnover cognitions. Journal of Vocational Behavior, 53: 254-273. Schwab, D. (1991) Contextual variables in employee performance-turnover relationships. Academy of Management Joumal, 34: 966-975.

International Journal of Manaj^cment

Vol. 24 No. 2

June 2007

229

Simons, T, & Hinkin, T. (2001) The effect of employee turnover on hotel profits: A test across multiple hotels. The Cornell Hotel and Restaurant Administration Quarterly, 42: 65-69. Somers, M.J. (1995) Organizational commitment, turnover, and absenteeism: An examination of direct and interaction effects. Journal of Organizational Behavior, 16:49-58. . - I Steers, R.M., & Mowday, R.T. (1981). Employee turnover and post decision accommodation processes. Pp. 235-281 in L.L. Cummings, & B.M. Staw (Eds.), Research in Organizational Behavior, Vol. 3. Greenwich, CT: JAI Press. Slopher, P.R., Meyburg, A.H., & Brg, W. (1981). New horizons in travel behavior research. Lexington. MA: Lexington Books. Taiwan Govemment Information Office (2004). Taiwan yearbook 2003. Taipei, Taiwan: Government Printing Office. Tseng. F.M., & Lin, L. (2005) A quadratic interval logit model for forecasting bankruptcy. Omega, 33:85-91. Tseng, F.M., & Yu, C.Y. (2005) Partitioned fuzzy integral multinomial logit model for Taiwan's internet telephony market. Omega, 33: 267-276. Trevor, C.,Gerhart, B., & Boundreau, J. (1997)Voluntary tumover and Job performance: Curviiincariiy and (he moderating influences of salary growth and promotions. Joumal of Applied Psychology, 82: 44-61. van Breukelen, W., van der Vlist, R., & Steensma, H. (2004) Voluntary employee turnover: combining variables from the 'traditional' tumover literature witb the theory of planned behavior. Journal of Organizational Behavior., 25: 893-915. Vecchio, R., & Norris, W. (1996) Voluntary tumover and job performance, satisfaction, and leader-member exchange. Journal of Business and Psychology, 11: 113-125. Williams, C.R., & Livingstone, L.P. (1994) Another look at the relationship between performance and voluntary tumover. Academy of Management Joumal, 37: 269-298. Contact email addresses: st6114.mei@msa.hinet.net y fchcn@mail.dyu.edu.tw noahsy.wei@msa.hinet.net